Hyun Dong Lee @ COLM2026(@hyundongleee):SwiLA's choices: ① Function class: a mixture of J linear regressions; each value coordinate picks one, with a learned prior. ② Optimizer: online SGD on the likelihood-based loss; each regressor takes a delta-rule step weighted by its responsibility. J = 1 recovers DeltaNet.

Hyun Dong Lee @ COLM2026

@hyundongleee

CS PhD student @Stanford

Joined July 2020

791 Following 201 Followers

Hyun Dong Lee @ COLM2026@hyundongleee

2026.10.05 16:42

SwiLA's choices: ① Function class: a mixture of J linear regressions; each value coordinate picks one, with a learned prior. ② Optimizer: online SGD on the likelihood-based loss; each regressor takes a delta-rule step weighted by its responsibility. J = 1 recovers DeltaNet.